I wouldn’t call it an air gap, but when you use a model on Bedrock or Vertex the inference runs on Amazon or Google’s servers respectively. The model provider gives them the weights, and is not otherwise in the loop for serving individual requests.
The harness was censored overnight as well. Can't ask it who won the 2020 election or who was the president in 2020 or 2000. It will say who is president right now though. Crazy shit really.
Often not, vendors like OpenAI or OpenAI through Azure might retain messages for safety reasons if they activate some safety filter, but it won't use it for training.
What about the model? My understanding is those queries are saved by the model provider.
Given this administration, and the privacy issues with the official WH app, this feels very misleading.